1. Patchew
  2. linux
  3. View series

[PATCH rdma-next] RDMA/bnxt_re: Clear VM_MAYWRITE on read-only mmap of driver pages

Leon Romanovsky posted 1 patch 2 days, 9 hours ago 
drivers/infiniband/hw/bnxt_re/ib_verbs.c | 13 ++++++++-----
1 file changed, 8 insertions(+), 5 deletions(-)
[PATCH rdma-next] RDMA/bnxt_re: Clear VM_MAYWRITE on read-only mmap of driver pages
Posted by Leon Romanovsky 2 days, 9 hours ago 
From: Leon Romanovsky <leonro@nvidia.com>

bnxt_re_mmap() rejects an initially writable mapping of the DBR pacing page
and the toggle page, but leaves VM_MAYWRITE set on the accepted read-only
mapping. A later mprotect(PROT_READ | PROT_WRITE) therefore passes the mm
permission check and upgrades the inserted PTEs, letting userspace write
these driver-owned pages: the DBR pacing parameters maintained under
rdev->pacing.dbq_lock, and the CQ/SRQ toggle state written from the NQ
tasklet.

Clear VM_MAYWRITE before vm_insert_page() so the mapping can never be made
writable, making any such mprotect() fail with -EACCES while the read-only
mapping continues to work.

Fixes: ea222485788208 ("RDMA/bnxt_re: Update alloc_page uapi for pacing")
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 drivers/infiniband/hw/bnxt_re/ib_verbs.c | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.c b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
index adc693736769..dcfb1b0ebc22 100644
--- a/drivers/infiniband/hw/bnxt_re/ib_verbs.c
+++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
@@ -4984,12 +4984,15 @@ int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
 		break;
 	case BNXT_RE_MMAP_DBR_PAGE:
 	case BNXT_RE_MMAP_TOGGLE_PAGE:
-		/* Driver doesn't expect write access for user space */
-		if (vma->vm_flags & VM_WRITE)
+		/* Reject writable mappings and prevent mprotect() upgrades. */
+		if (vma->vm_flags & VM_WRITE) {
 			ret = -EFAULT;
-		else
-			ret = vm_insert_page(vma, vma->vm_start,
-					     virt_to_page((void *)bnxt_entry->mem_offset));
+			break;
+		}
+
+		vm_flags_clear(vma, VM_MAYWRITE);
+		ret = vm_insert_page(vma, vma->vm_start,
+				     virt_to_page((void *)bnxt_entry->mem_offset));
 		break;
 	default:
 		ret = -EINVAL;

---
base-commit: 0e8e94c15091041ea8910cbfcade5a9c7cfe3f90
change-id: 20260722-missing-vma-write-protection-enforce-2e99624c35dc

Best regards,
--  
Leon Romanovsky <leonro@nvidia.com>
Re: [PATCH rdma-next] RDMA/bnxt_re: Clear VM_MAYWRITE on read-only mmap of driver pages
Posted by Selvin Xavier 2 days, 8 hours ago 
On Wed, Jul 22, 2026 at 6:26 PM Leon Romanovsky <leon@kernel.org> wrote:
>
> From: Leon Romanovsky <leonro@nvidia.com>
>
> bnxt_re_mmap() rejects an initially writable mapping of the DBR pacing page
> and the toggle page, but leaves VM_MAYWRITE set on the accepted read-only
> mapping. A later mprotect(PROT_READ | PROT_WRITE) therefore passes the mm
> permission check and upgrades the inserted PTEs, letting userspace write
> these driver-owned pages: the DBR pacing parameters maintained under
> rdev->pacing.dbq_lock, and the CQ/SRQ toggle state written from the NQ
> tasklet.
>
> Clear VM_MAYWRITE before vm_insert_page() so the mapping can never be made
> writable, making any such mprotect() fail with -EACCES while the read-only
> mapping continues to work.
>
> Fixes: ea222485788208 ("RDMA/bnxt_re: Update alloc_page uapi for pacing")
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> ---
>  drivers/infiniband/hw/bnxt_re/ib_verbs.c | 13 ++++++++-----
>  1 file changed, 8 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.c b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
> index adc693736769..dcfb1b0ebc22 100644
> --- a/drivers/infiniband/hw/bnxt_re/ib_verbs.c
> +++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
> @@ -4984,12 +4984,15 @@ int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
>                 break;
>         case BNXT_RE_MMAP_DBR_PAGE:
>         case BNXT_RE_MMAP_TOGGLE_PAGE:
> -               /* Driver doesn't expect write access for user space */
> -               if (vma->vm_flags & VM_WRITE)
> +               /* Reject writable mappings and prevent mprotect() upgrades. */
> +               if (vma->vm_flags & VM_WRITE) {
>                         ret = -EFAULT;
> -               else
> -                       ret = vm_insert_page(vma, vma->vm_start,
> -                                            virt_to_page((void *)bnxt_entry->mem_offset));
> +                       break;
> +               }
> +
> +               vm_flags_clear(vma, VM_MAYWRITE);
> +               ret = vm_insert_page(vma, vma->vm_start,
> +                                    virt_to_page((void *)bnxt_entry->mem_offset));
Same change was done as part of my series which cleaned up the toggle
page for CQ/SRQ.
https://lore.kernel.org/linux-rdma/20260721115440.24021-5-selvin.xavier@broadcom.com/
Do you want to take this and abandon my patch? I am okay with that. I
can rebase my series
once this gets merged.

Thanks,
Selvin

>                 break;
>         default:
>                 ret = -EINVAL;
>
> ---
> base-commit: 0e8e94c15091041ea8910cbfcade5a9c7cfe3f90
> change-id: 20260722-missing-vma-write-protection-enforce-2e99624c35dc
>
> Best regards,
> --
> Leon Romanovsky <leonro@nvidia.com>
>
Re: [PATCH rdma-next] RDMA/bnxt_re: Clear VM_MAYWRITE on read-only mmap of driver pages
Posted by Leon Romanovsky 2 days, 6 hours ago 
On Wed, Jul 22, 2026 at 07:05:52PM +0530, Selvin Xavier wrote:
> On Wed, Jul 22, 2026 at 6:26 PM Leon Romanovsky <leon@kernel.org> wrote:
> >
> > From: Leon Romanovsky <leonro@nvidia.com>
> >
> > bnxt_re_mmap() rejects an initially writable mapping of the DBR pacing page
> > and the toggle page, but leaves VM_MAYWRITE set on the accepted read-only
> > mapping. A later mprotect(PROT_READ | PROT_WRITE) therefore passes the mm
> > permission check and upgrades the inserted PTEs, letting userspace write
> > these driver-owned pages: the DBR pacing parameters maintained under
> > rdev->pacing.dbq_lock, and the CQ/SRQ toggle state written from the NQ
> > tasklet.
> >
> > Clear VM_MAYWRITE before vm_insert_page() so the mapping can never be made
> > writable, making any such mprotect() fail with -EACCES while the read-only
> > mapping continues to work.
> >
> > Fixes: ea222485788208 ("RDMA/bnxt_re: Update alloc_page uapi for pacing")
> > Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> > ---
> >  drivers/infiniband/hw/bnxt_re/ib_verbs.c | 13 ++++++++-----
> >  1 file changed, 8 insertions(+), 5 deletions(-)
> >
> > diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.c b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
> > index adc693736769..dcfb1b0ebc22 100644
> > --- a/drivers/infiniband/hw/bnxt_re/ib_verbs.c
> > +++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
> > @@ -4984,12 +4984,15 @@ int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
> >                 break;
> >         case BNXT_RE_MMAP_DBR_PAGE:
> >         case BNXT_RE_MMAP_TOGGLE_PAGE:
> > -               /* Driver doesn't expect write access for user space */
> > -               if (vma->vm_flags & VM_WRITE)
> > +               /* Reject writable mappings and prevent mprotect() upgrades. */
> > +               if (vma->vm_flags & VM_WRITE) {
> >                         ret = -EFAULT;
> > -               else
> > -                       ret = vm_insert_page(vma, vma->vm_start,
> > -                                            virt_to_page((void *)bnxt_entry->mem_offset));
> > +                       break;
> > +               }
> > +
> > +               vm_flags_clear(vma, VM_MAYWRITE);
> > +               ret = vm_insert_page(vma, vma->vm_start,
> > +                                    virt_to_page((void *)bnxt_entry->mem_offset));
> Same change was done as part of my series which cleaned up the toggle
> page for CQ/SRQ.
> https://lore.kernel.org/linux-rdma/20260721115440.24021-5-selvin.xavier@broadcom.com/
> Do you want to take this and abandon my patch? I am okay with that. I
> can rebase my series
> once this gets merged.

I took your patch as you posted before me and added Fixes line.

Thanks

> 
> Thanks,
> Selvin
> 
> >                 break;
> >         default:
> >                 ret = -EINVAL;
> >
> > ---
> > base-commit: 0e8e94c15091041ea8910cbfcade5a9c7cfe3f90
> > change-id: 20260722-missing-vma-write-protection-enforce-2e99624c35dc
> >
> > Best regards,
> > --
> > Leon Romanovsky <leonro@nvidia.com>
> >

Patchew 2.1-devel-ea86937

© 2016 - 2026 Red Hat, Inc.